For our analyzed study group, cases compared to controls were more likely to be men, be slightly older, and have ever smoked (and if a past smoker, had tended to have quit more recently). Cases were more likely to have DM (especially new onset; columns 1 and 2) compared to controls. Usual BMI was similar between cases and controls. When comparing male and female cases (not shown), men were slightly younger, had a higher usual BMI, were more likely to have ever smoked (difference seen mostly among ex-smokers rather than current smokers), and were more likely to have DM (especially if diagnosed greater than 3 years before pancreatic adenocarcinoma).
Characteristics of pancreatic adenocarcinoma cases and controls used in the analysis (no recent diet change), and all adenocarcinoma cases and controls who completed FFQs
Comparing the demographic characteristics of our analyzed participants to those who completed a FFQ but who were excluded showed that the two groups were similar. The data available to us on all of our cases permitted us to assess the potential effect of not including data on the 1,549 patients who returned FFQs. A comparison of demographic characteristics for the 384 analyzed cases versus all 1,933 cases who completed a FFQ showed that our analyzed cases were more likely to be men, older, have DM (especially onset < 3 years prior to study), and be a current smoker. The 983 analyzed controls and all 1,713 controls who completed a FFQ were very similar for most demographic characteristics, with the largest deviations occurring for DM and smoking (especially former smokers who quit < 10 years prior to study), with the analyzed participants having lower values.
The median value and interquartile range are given for each food grouping (e.g., fruits, vegetables, fibers, and grains) by sex and case status in . We noted that the female controls have the highest intake value for most of the groupings and male cases have the lowest value. In general, female cases and male controls have similar values for each food grouping, with female cases having slightly higher median values. These patterns justified reporting of median sex-specific dietary intake and constructing quintiles separately by sex for our sample.
Food groupings analyzed among pancreatic adenocarcinoma cases and controls
Many of the food groupings in our study are moderately (Pearson’s r2 60–80) or highly correlated (Pearson’s r2 ≥ 0.80). The “total fruit” grouping is highly correlated with “citrus, melon, and berries” (0.85) and “other fruit” (0.80) groupings. The “total vegetable” grouping is highly correlated with “other vegetable” (0.81) grouping and moderately correlated with “dark green vegetable” (0.62) grouping. “Total dietary fiber” is highly correlated with “soluble dietary fiber” (0.88) and “insoluble dietary fiber” (0.98) groupings. “Total grain” is highly correlated with “non-whole grains” (0.83), and moderately correlated with “whole grains” (0.69).
also shows the number of cases along with an OR and 95% CI, and trend test of OR across quintiles, which were constructed using the control population values for each sex. Associations between food groupings and pancreatic adenocarcinoma were determined by using a multivariable logistic regression model adjusting for age, sex, energy (per 1,000 kcal), BMI, smoking (current, former, never), and drinks of alcohol per week and compared the four highest quintiles to the reference (lowest consumption). Significant results (test for trend p-value (ptrend) < 0.05) for an inverse association between pancreatic adenocarcinoma and food groupings (OR [95% CI]) were citrus, melon, and berries (OR = 0.70[0.47–1.04]), other fruit (OR = 0.73[0.49–1.10]), total fruit (OR = 0.57[0.37–0.86]), dark green vegetable (OR = 0.43 [0.28–0.65]), deep yellow vegetable (OR = 0.58[0.39–0.86]), tomato (OR = 0.57[0.38–0.86]), other vegetable (OR = 0.49[0.32–0.75]), dry bean and pea (OR = 0.81[0.55–1.20]), total vegetables (OR = 0.56[0.37–0.84]), insoluble fiber (OR = 0.48[0.33–0.71]), soluble fiber (OR = 0.58[0.39–0.86]), total dietary fiber (OR = 0.47[0.32–0.70]), whole grains (OR = 0.70[0.47–1.03]) and orange/grapefruit juice (OR = 0.52[0.35–0.79]). There was an increased association between having pancreatic adenocarcinoma and non-whole grains (OR = 2.10[1.38–3.20]). It is important to note that the correlation between whole and non-whole grains was low (r2 = 0.17); therefore, the discordant associations for these categories are not likely due to simple dietary replacement of one grain for the other.
Because DM is a known risk factor for pancreatic adenocarcinoma and because diabetics may be advised to modify their diet, we investigated whether adding DM (categorized as no DM, DM diagnosis less than 3 years prior to completing questionnaire, or DM diagnosis 3 years or more prior to questionnaire) to our logistic model would significantly change our results (not shown). There were two food groupings where higher consumption was inversely associated with having pancreatic adenocarcinoma, and that changed in statistical significance after adding DM to the analyses (“citrus, melon, and berries”: ptrend changed from 0.03 to 0.07 and fifth (highest intake) quintile compared to first (lowest intake) quintile OR (95% CI) changed from 0.70 (0.47–1.04) to 0.75 (0.49–1.17); and “other starches”: ptrend changed from 0.21 to 0.05 and fifth quintile compared to first quintile OR (95% CI) changed from 0.92 (0.62–1.35) to 0.76 (0.49–1.19). We also examined the presented model in only non-diabetics. There were two significant differences between the presented model and the non-DM model analysis that were not significantly different when DM adjustment was used (for “dry bean and pea” ptrend changed from 0.03 to 0.32) and fifth quintile compared to first quintile OR (95% CI) changed from 0.81 (0.55–1.20) to 0.84 (0.50–1.39), and for “tomatoes,” ptrend changed from 0.003 to 0.24 and fifth quintile compared to first quintile OR (95% CI) changed from 0.57 (0.38–0.86) to 0.68 (0.41–1.12), indicating a possible interaction effect which is beyond the scope of this study to address further.
In addition, total sugar intake was considered as a possible confounder for our fruit and vegetable food groupings and was adjusted for in an alternative version of the presented model. Higher consumption of one food grouping was inversely associated with having pancreatic adenocarcinoma, which became statistically non-significant when adjusted for sugar intake in the analyses (“dry bean and pea” ptrend changed from 0.03 to 0.08).
A sensitivity analysis was conducted (data not shown) to determine the impact of failure to exclude those who reported a diet change within the last 5 years. There was one food grouping (associated with an increased risk of pancreatic adenocarcinoma) that increased in statistical significance when those reporting a recent diet change were included in the analyses (“other starches,” ptrend changed from 0.22 to 0.03). Among males who reported a recent diet change, the median values for both fruit groupings and both fiber groupings decreased. Among women, the median values increased for both fruit groupings and decreased for both the fiber and grain groupings. These observed changes affirm the importance of our strategy to use the recent diet change data to exclude subjects when determining our final sample, increasing confidence that our associations may be related to causation.
To investigate possible residual confounding by smoking, we conducted a sensitivity analysis restricting our data to only never smokers (). The categories of citrus, melon, and berries (OR = 0.70[95% CI 0.47–1.04]), other fruit (OR = 0.73[95% CI 0.49–1.10]), dry bean and pea (OR = 0.81[95% CI 0.55–1.20]), and whole grains (OR = 0.70[95% CI 0.47–1.03]) were no longer significant in the subanalysis. However, only the OR for the highest quintile of the citrus, melon, and berries category changed by over 10%, suggesting that the significant changes were related to a reduced sample size. As a finer stratification of smoking exposure, we used a pack-years categorical adjustment in our logistic regression model and found no significant difference from the presented model results () for associations between any food groupings and pancreatic adenocarcinoma.
Food groupings associated with pancreatic adenocarcinoma among never smokers
The demographic characteristic comparisons of participants who did and did not complete the questionnaire after recruitment showed that the two groups were similar. The data available to us on all of our cases permitted us to assess the potential effect of not capturing data on the 832 patients who did not return FFQs. The demographic characteristics of the 816 patients who completed the FFQ and the 832 patients who did not were respectively ~9 months versus ~6 months for median days from approach to death. The median age for patients not completing the FFQ was 65.9 years, compared to 66.3 for those who did. When grouped by stage, 31.3% of early-stage (I and II) patients completed the FFQ, compared to 44.7% of late stage (III and IV). There were also modest differences among those who completed the FFQ compared to those who did not by usual BMI (median, 27.5 vs. 28.4), DM (38.1% vs. 49%), and current smoker (23.6 vs. 15.6%) and former smoker status (36.9 vs. 43.4%). We excluded participants who reported a recent change in diet (420 of the 816 cases), and we predict that a similar or higher proportion of the 832 patients who did not complete the FFQ would have been excluded if we had their data. Moreover, the impact of smoking and BMI would have been adjusted in the model, and we anticipate that DM would not have a large effect on results based on our analyzed dataset.